The Agent Skills Directory

[METADATA_POISONING]: The skill metadata identifies the author as 'harness-engineering', which conflicts with the provided context identifying the author as 'kostja94'. This discrepancy is deceptive regarding the skill's origin.
[COMMAND_EXECUTION]: The skill performs shell-based network probing using curl, nslookup, and Invoke-WebRequest. These tools are used for their intended purposes (fetching headers and DNS records) and are correctly declared in the allowed-tools metadata field.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from remote websites across multiple layers (L3 robots.txt, L5 HTML source, L7 network captures).
Ingestion points: Content is fetched via WebFetch, curl, and browser automation in SKILL.md.
Boundary markers: The instructions lack explicit boundary markers or warnings to the agent to ignore embedded instructions in the fetched data.
Capability inventory: The skill has access to shell environments and browser automation (browser_evaluate, browser_network).
Sanitization: There is no evidence of sanitization for prompt injection patterns within the fetched content. The risk is partially mitigated by the skill's heavy reliance on deterministic signature matching against local reference files.
[DYNAMIC_EXECUTION]: The skill uses browser_evaluate to run a locally bundled JavaScript file (scripts/detect-js-globals.js). The script performs passive inspection of the DOM and global window objects to detect framework versions and service markers, which is a standard and benign use of dynamic execution for this use case.

tech-stack-detect